From 7969dc48000ce1c2fb25fdc6b07865ca52bfede0 Mon Sep 17 00:00:00 2001
From: Ian Barwick <ian@2ndquadrant.com>
Date: Thu, 8 Oct 2020 11:28:10 +0900
Subject: [PATCH] Enable "node rejoin" to join a target with a lower timeline

This has been possible since PostgreSQL 9.6, but the node rejoin/follow
check did not consider this possibility.
---
 HISTORY                        |  1 +
 doc/appendix-release-notes.xml |  8 +++++
 doc/repmgr-node-rejoin.xml     |  5 +++
 repmgr-client.c                | 64 ++++++++++++++++++++++++++++------
 4 files changed, 68 insertions(+), 10 deletions(-)

diff --git a/HISTORY b/HISTORY
index 225756d2..a5de2e82 100644
--- a/HISTORY
+++ b/HISTORY
@@ -11,6 +11,7 @@
         repmgr: report database connection error if the --optformat option was
           provided to "node check" (Ian)
         repmgr: improve "node rejoin" checks (Ian)
+        repmgr: enable "node rejoin" to join a target with a lower timeline (Ian)
         repmgr: improve output formatting for cluster matrix/crosscheck (Ian)
         repmgr: improve database connection failure error checking on the
           demotion candidate during "standby switchover" (Ian)
diff --git a/doc/appendix-release-notes.xml b/doc/appendix-release-notes.xml
index 03fb890d..00589481 100644
--- a/doc/appendix-release-notes.xml
+++ b/doc/appendix-release-notes.xml
@@ -107,6 +107,14 @@
             </para>
           </listitem>
 
+          <listitem>
+            <para>
+              <link linkend="repmgr-node-rejoin">repmgr node rejoin</link>:
+              enable a node to attach to a target node even the target node
+              has a lower timeline (PostgreSQL 9.6 and later).
+            </para>
+          </listitem>
+
           <listitem>
             <para>
               <link linkend="repmgr-node-check">repmgr node check</link>:
diff --git a/doc/repmgr-node-rejoin.xml b/doc/repmgr-node-rejoin.xml
index 0427e422..8500db58 100644
--- a/doc/repmgr-node-rejoin.xml
+++ b/doc/repmgr-node-rejoin.xml
@@ -392,6 +392,11 @@
      the current standby's PostgreSQL log will contain entries with the text
      &quot;<literal>record with incorrect prev-link</literal>&quot;.
    </para>
+   <para>
+     In PostgreSQL 9.5 and earlier, it is <emphasis>not</emphasis> possible to use
+     <application>pg_rewind</application> to attach to a target node with a lower
+     timeline than the local node.
+   </para>
    <para>
      We strongly recommend running <command>repmgr node rejoin</command> with the
      <option>--dry-run</option> option first. Additionally it might be a good idea
diff --git a/repmgr-client.c b/repmgr-client.c
index 737e6fa8..942dd46e 100644
--- a/repmgr-client.c
+++ b/repmgr-client.c
@@ -4044,7 +4044,7 @@ check_standby_join(PGconn *upstream_conn, t_node_info *upstream_node_record, t_n
 
 /*
  * Here we'll perform some timeline sanity checks to ensure the follow target
- * can actually be followed.
+ * can actually be followed or rejoined.
  *
  * See also comment for check_node_can_follow() in repmgrd-physical.c .
  */
@@ -4130,20 +4130,64 @@ check_node_can_attach(TimeLineID local_tli, XLogRecPtr local_xlogpos, PGconn *fo
 				action,
 				follow_target_identification.timeline);
 
-	/* upstream's timeline is lower than ours - impossible case */
+	/*
+	 * The upstream's timeline is lower than ours - we cannot follow, and rejoin
+	 * requires PostgreSQL 9.6 and later.
+	 */
 	if (follow_target_identification.timeline < local_tli)
 	{
-		log_error(_("this node's timeline is ahead of the %s target node's timeline"), action);
-		log_detail(_("this node's timeline is %i, %s target node's timeline is %i"),
-				   local_tli,
-				   action,
-				   follow_target_identification.timeline);
-		PQfinish(follow_target_repl_conn);
-		return false;
+		/*
+		 * "repmgr standby follow" is impossible in this case
+		 */
+		if (is_rejoin == false)
+		{
+			log_error(_("this node's timeline is ahead of the %s target node's timeline"), action);
+			log_detail(_("this node's timeline is %i, %s target node's timeline is %i"),
+					   local_tli,
+					   action,
+					   follow_target_identification.timeline);
+
+			if (PQserverVersion(follow_target_conn) >= 90600)
+			{
+				log_hint(_("use \"repmgr node rejoin --force-rewind\" to reattach this node"));
+			}
+
+			PQfinish(follow_target_repl_conn);
+			return false;
+		}
+
+		/*
+		 * pg_rewind can only rejoin to a lower timeline from PostgreSQL 9.6
+		 */
+		if (PQserverVersion(follow_target_conn) < 90600)
+		{
+			log_error(_("this node's timeline is ahead of the %s target node's timeline"), action);
+			log_detail(_("this node's timeline is %i, %s target node's timeline is %i"),
+					   local_tli,
+					   action,
+					   follow_target_identification.timeline);
+
+			if (runtime_options.force_rewind_used == true)
+			{
+				log_hint(_("pg_rewind can only be used to rejoin to a node with a lower timeline from PostgreSQL 9.6"));
+			}
+
+			PQfinish(follow_target_repl_conn);
+			return false;
+		}
+
+		if (runtime_options.force_rewind_used == false)
+		{
+			log_notice(_("pg_rewind execution required for this node to attach to rejoin target node %i"),
+					   follow_target_node_record->node_id);
+			log_hint(_("provide --force-rewind"));
+			PQfinish(follow_target_repl_conn);
+			return false;
+		}
 	}
 
 	/* timelines are the same - check relative positions */
-	if (follow_target_identification.timeline == local_tli)
+	else if (follow_target_identification.timeline == local_tli)
 	{
 		XLogRecPtr follow_target_xlogpos = get_node_current_lsn(follow_target_conn);